Easy methods to Use MCP: Mannequin Context Protocol

You’ve constructed functions with LLMs. You’ve performed with brokers. Possibly you’ve even labored with LangChain, AutoGen, or OpenAI’s Assistants API. Isn’t it spectacular how a lot these fashions can motive, perceive, and generate? However the second your agent must do one thing actual, like examine a database, learn from a CRM, or fetch a Google Doc; issues begin to disintegrate. Immediately you’re:

• Writing customized wrappers for every API,
• Manually deciding what information goes within the immediate, and
• Combating token limits and off context.

That’s the place Mannequin Context Protocol (MCP) is available in. MCP shouldn’t be a brand new “framework” however an open commonplace that helps AI assistants and brokers hook up with real-world information and instruments in a easy and reusable approach. On this weblog, we’ll perceive what MCP is, use it, and why it’s so well-liked now – all whereas exploring its functions.

What’s MCP?

Mannequin Context Protocol (MCP) is a robust open commonplace that was launched by Claude’s guardian firm, Anthropic, in November 2024. It permits Giant Language Fashions (LLMs) like Claude or GPT to seamlessly entry exterior content material equivalent to paperwork, databases, instruments, native information or APIs, in a structured, safe, and scalable approach.

MCP at its core consists of three primary elements:

1. Mannequin: It refers back to the LLM (like Claude or GPT-4) that’s responding to your inputs.
2. Context: It’s the extra data the mannequin wants to reply meaningfully. The supply of this data will be paperwork, PDFs, prompts, or database entries.
3. Protocol: It represents a algorithm that lets the mannequin entry and use that context from structured sources.

How Does MCP Work?

In its true essence, MCP standardizes the communication between the purchasers and servers, without having customized code or handbook information uploads. Shoppers and Servers can imply various things for various applied sciences, however for MCP, these are:. 

• Shopper: That is the interface the place the LLMs reply to you. This could possibly be Claude’s desktop app, your IDE, or a chatbot.
• Server: That is the system or information supply (native or distant) that holds the context, for eg. Google Drive, GitHub, Gmail, your native information, PDFs, and so on.

If you ask a query or give a command, the shopper sends a request to the server for the related data. The server then offers the required context (e.g., a file, electronic mail, or database), which the shopper makes use of to present you a response or full a process. This course of permits AI to effectively work with real-time, dynamic information.

Thus, with MCP, the shopper (LLM interface) can robotically uncover obtainable servers (information sources), and request particular data primarily based in your question.

Why Do We Want MCP?

MCP acts like a common connector to your LLMs, permitting them to:

• Discover out what instruments can be found (like your information, calendar, or paperwork).
• Perceive what every instrument can do (e.g. “learn PDF”, “depend information”, “save chat”).
• Use these instruments instantly without having any particular code for every one.
• Get clear, structured outcomes which it could possibly use to reply you higher.

So as an alternative of you importing information or copying information into the immediate, the LLMs simply entry what it wants robotically. Whether or not it’s studying from a Notion doc, querying a database, or retrieving a GitHub difficulty, MCP turns it right into a seamless, reusable workflow.

Easy methods to Use MCP?

On this tutorial, we’ll use MCP to attach the Claude desktop app (shopper) to our native information and information utilizing Python. We’ll then use the exterior information as context to carry out these 4 duties:

1. Create a customized greeting instrument
2. Depend the variety of information on our desktop
3. Save a chat with Claude to our native system
4. Ask Claude questions utilizing an area PDF doc

Necessities

To carry out all these 4 duties, just be sure you have:

1. A system with Python 3.11+ put in
2. Claude desktop app (free or professional model)
3. Fundamental familiarity with Python
4. VS Code or any IDE
5. Just a few native paperwork (e.g. a PDF like a go away coverage)

One-time Set Up

Earlier than we begin with performing the duties, listed below are a couple of one time duties that it is advisable to carry out to make sure all of your 4 duties are carried out easily. 

Create a Python Atmosphere

It’s extremely really helpful to create a separate python atmosphere to carry out these duties. If you’re utilizing a conda atmosphere, you are able to do that by executing the next command:

conda create ~n mcp python==3.11

Set up the MCP Library

Since we’re working with Python, it’s necessary to include Python code information. To do that, write  and execute the command:

pip set up mcp

The code file, in addition to Python will also be in Typescript, Java or Kotlin. You’ll be able to learn extra about these in MCP Documentation.

Obtain Claude’s Desktop App

To obtain the app, head to https://claude.ai/obtain and obtain the app to your respective system. As soon as put in, it could take a couple of seconds to launch. It’s utterly regular. Though, yoy can construct with Claude MCP utilizing Claude’s free plan however the plan its utilization limits whereas experimenting and is commonly inconvenient to work with. Therefore its most well-liked to work with a paid subscription.

Set-up the Configuration File

The configuration file is exclusive for every shopper. Since we’re utilizing Claude’s desktop app, we might want to obtain Claude’s configuration file. To do that:

• Go to Claude Settings.
• Then go to the Developer Part.
• Click on on “Edit Configuration”.



• Search for “claude_desktop_config.json”. If the file isn’t already there, please create it with the identical title.



This file maps your instruments (Python scripts) to Claude and also you’ll replace this file after every process. Now that we’re accomplished with all our installations and downloads, let’s begin with our duties. 

Activity 1: Create a Customized Greeting Software

On this process, we’ll construct a instrument that can create a personalised greeting for you every time prompted with a “Hello” or “Hiya”.

Step 1: Create a New Python File

Right here, I’ve created a brand new file named “greeter.py”.

from mcp.server.fastmcp import FastMCP// FAST MCP is simply a neater method to construct a server

# Create an MCP server named "Greeter"

mcp = FastMCP("Greeter") 

@mcp.instrument()

def greet() -> str:

    """Return this welcome message, when greeted with "Hello", "Hey" or "Hiya"."""

    return "Hey <add your personal title>, Welcome to the world of MCPs!"

if __name__ == "__main__":

    mcp.run()

Code Rationalization:

The code creates an MCP server that reads for a greeting request and responds with a personalised message. 

• The perform makes use of a decorator to inform the server reply when prompted with “Hello“, “Hey“, or “Hiya“. 
• The command we give to the MCP server will get shared with the LLM, which in flip would possibly generate the identical response we fed to the server or a model just like it. 
• You could get a very completely different response than what you count on; it’s because the ultimate output is generated by an LLM.

Step 2: Replace the Config File

Be certain there are completely no syntax errors in your configuration file. In a configuration file we want three issues:

1. Identify of our server
2. Location of the Python program
3. Location of the Python file that must be executed



Step 3: Restart Your Claude Desktop App

Restarting the app is important as a result of,

1. It tells you about any errors in your code. The error will get highlighted and you’ll go to the “Logs” part to get insights on how these errors will be corrected.
2. When no errors are detected, the instrument will get registered to the Claude desktop app and it is possible for you to to entry it from inside the app.
3. Restarting the app provides it entry to the newest integrations. So it is advisable to do it each time you replace or change your code.

Step 4: Say “Hello” to Claude within the App

Simply immediate a “Hello” to Claude through the app and it’ll greet you again utilizing your instrument!

When you say “hey” to the Claude desktop app, you’re going to get a pop-up asking you to permit the app to make use of the instrument that you’ve got simply created. Click on on “Enable instrument for this chat”. At the moment it’s important to enable a instrument to work in a chat each time you name that instrument in a given chat.



As soon as permission is given, you get the response from the desktop app. 



Activity 2: Depend Information on Your Desktop

This instrument will depend the variety of information current in your desktop and return the overall whenever you ask Claude.

Step 1: Create a New Python File

Right here, I’m making a file named file_counter.py

from mcp.server.fastmcp import FastMCP

import os

# Create an MCP server named "FileCounter"

mcp = FastMCP("FileCounter")

@mcp.instrument()

def count_desktop_files() -> str:

    """Depend the variety of information on the desktop"""

    desktop_path = os.path.expanduser("~/Desktop")  # Will get desktop path (e.g., /Customers/apoorv/Desktop)

    attempt:

        # Checklist all objects in desktop listing, filter to information solely

        information = [f for f in os.listdir(desktop_path) if os.path.isfile(os.path.join(desktop_path, f))]

        file_count = len(information)

        return f"There are {file_count} information in your desktop."

    besides Exception as e:

        return f"Error counting information: {str(e)}"

if __name__ == "__main__":

    mcp.run()

Code Rationalization:

This code units up an MCP server to scan the desktop listing, depend the variety of information, and return the depend. The perform filters via the listing utilizing Python’s os library to depend solely the information (not folders).

Step 2: Replace the Config File

Comply with the steps as above and ensure there aren’t any additional information. 

Step 3: Restart Your Claude Desktop App

Comply with the steps as above. 

Step 4: Ask Claude

Give the next immediate to the Claude Desktop app: “What number of information are there on my desktop?”

Then you’re going to get the pop as much as enable the instrument for this chat. As soon as permitted you’re going to get the output primarily based on the variety of information in your system.



Activity 3: Save Conversations as Textual content Information

This instrument permits Claude to save lots of your ongoing dialog right into a .txt file in your desktop.

Step 1: Create a New Python File

For this, I’m going to create a brand new file named conversation_saver.py.

from mcp.server.fastmcp import FastMCP

import os

from datetime import datetime 

mcp = FastMCP("ConversationSaver")

@mcp.instrument()

def save_conversation(dialog: str) -> str:

    """Save the present dialog to a textual content file on the desktop"""

    desktop_path = os.path.expanduser("~/Desktop")

    timestamp = datetime.now().strftime("%Ypercentmpercentd_percentHpercentMpercentS")

    file_path = os.path.be a part of(desktop_path, f"claude_chat_{timestamp}.txt")

    attempt:

        with open(file_path, "w") as f:

            f.write(dialog)

        return f"Dialog saved at: {file_path}"

    besides Exception as e:

        return f"Error saving dialog: {str(e)}"

if __name__ == "__main__":

    mcp.run()

Code Rationalization:

The code creates an MCP server that saves the continuing dialog to a textual content file in your desktop, naming it with a timestamp. It makes use of Python’s datetime to generate distinctive file names and the os library to save lots of the content material in a textual content file.

Step 2: Replace the Config File

Be certain there are completely no syntax errors in your configuration file and there aren’t any additional information within the record. 

Step 3: Restart Your Claude Desktop App

To examine for any errors in your code, restart the app after writing the code. Any errors get highlighted after this step and you’ll appropriate them. If there aren’t any errors, your instrument will get registered to the Claude app, after which you’ll entry it instantly from there. 

Step 4: Save the Chat

I’ve gave a immediate to the Claude App: “What’s a RAG system”

This time, it’s NOT going to entry any instruments; since we haven’t to this point created any instrument for the sort of question. So the app received’t ask you for any permissions and would proceed like AI chatbots normally do. It’s going to generate an output primarily based on its coaching data.

Now I give the app, a second immediate: “Save this chat”

Claude will immediate you to permit the instrument after which save the total dialog to your desktop, timestamped.



You’ll be able to open the .txt file  in your desktop to view your complete alternate.

Examine the output I acquired right here.

Activity 4: Ask Questions A few Native PDF

This instrument lets Claude entry a doc (which might  be phrase doc, pdf or ppt) obtainable in your native system, and reply questions from it; with out you having to add it manually. For this process, I’ll be prompting some questions relating to my firm’s go away coverage; the pdf for which is current at my desktop.

Step 1: Create a New Python File

For this process, I’m going to create a brand new file named leave_policy_reader.py.

from mcp.server.fastmcp import FastMCP

import os

import PyPDF2  # For PDF textual content extraction

mcp = FastMCP("SmartContext")

# Hardcoded path to your go away coverage PDF (edit this to your file’s location)

LEAVE_POLICY_PDF = "/Customers/apoorv/Desktop/AV/Code/MCP/leave_policy.pdf"

@mcp.instrument()

def fetch_context(question: str) -> str:

    """Fetch content material from a go away coverage PDF for leave-related queries"""

    query_lower = question.decrease()

    # Solely course of if "go away" is within the question

    if "go away" not in query_lower:

        return "This instrument solely solutions leave-related questions. Ask about leaves!"

    if not os.path.exists(LEAVE_POLICY_PDF):

        return f"Error: Go away coverage PDF not discovered at {LEAVE_POLICY_PDF}!"

    attempt:

        # Learn the PDF

        with open(LEAVE_POLICY_PDF, "rb") as f:

            pdf_reader = PyPDF2.PdfReader(f)

            content material = ""

            for web page in pdf_reader.pages:

                extracted_text = web page.extract_text()

                if extracted_text:

                    content material += extracted_text + "n"

            if not content material:

                return "Error: No textual content could possibly be extracted from the go away coverage PDF!"

        # Return the content material with the question for Claude

        return (f"Utilizing context from {LEAVE_POLICY_PDF}:nn"

                f"Content material:n{content material}nn"

                f"Question: {question}nn"

                f"Please reply primarily based on the content material above.")

    besides Exception as e:

        return f"Error studying go away coverage PDF: {str(e)}"

if __name__ == "__main__":

    mcp.run()

Code Rationalization:

This code builds an MCP server that reads a PDF doc (go away coverage) and solutions leave-related queries. It makes use of PyPDF2 to extract textual content from the PDF and responds primarily based on the doc’s content material. It solely will get activated if the question is about “go away”.

Step 2: Replace the Config File

As has been accomplished within the earlier process.

Step 3: Restart Your Claude Desktop App

As has been accomplished within the earlier process.



Step 4: Ask Claude

To check this instrument, I added the next immediate to my Claude app: “Can I take half day go away?”

Claude processes the immediate, realises it wants the entry to the instrument we had created and prompts us to grant it entry to the instrument.

As soon as the instrument is known as, it will compile the data from the go away coverage doc and the unique question. It will then feed it to the claude LLM and ask it to reply primarily based on it. To do that, the instrument will learn your PDF in actual time and share the data to the LLM.

Lastly you get the response primarily based on the go away coverage.



This manner you’ll be able to construct customised instruments to entry completely different information in your system with out having to add the information repeatedly!

Benefits of Utilizing MCP

After the above hands-on, it’s simpler to reply this query. Utilizing mannequin context protocol:

• Eliminates handbook information uploads or customized integrations.
• Permits LLMs to work with each native and cloud-based methods.
• Standardises the way in which any shopper works with a server so long as they’re utilizing the identical protocol.
• Helps SDKs in Python, TypeScript, Kotlin, and Java.

What Are the Prospects with MCP?

MCP brings with itself a bag of alternatives and potentialities. Listed below are a couple of methods it could possibly improve present workflows:

1. Multi-Step Initiatives: WIth MCP, AI methods can seamlessly coordinate duties throughout a number of platforms, like planning occasions or managing tasks, with out advanced integrations.
2. Actual-World Consciousness: It permits AI to work together with good environments, IoT gadgets, and OS capabilities, making AI extra proactive and intuitive.
3. Collaborating Brokers: Utilizing MCP, a number of specialised AI brokers can work collectively, exchanging data and collaborating dynamically with out direct integrations.
4. Private AI Assistants: With MCP, deeply customized assistants will be created that securely entry private information and gadgets, with out exposing delicate data to 3rd events.
5. Enhanced Buyer Help: Mixed with MCP, AI methods can present context-aware customer support by accessing real-time information, bettering response high quality and effectivity.
6. Doc Processing: It permits AI fashions to learn, course of, and reply primarily based on paperwork saved domestically or within the cloud, equivalent to contracts or authorized insurance policies.

These are only some of the limitless potentialities of MCP throughout varied fields like software program growth, healthcare, finance and extra!

Why Are We Immediately Speaking About MCP?

To date, we now have lined every part about MCP – from what it’s to use it, why to make use of it, and its functions. However there’s nonetheless one query that must be answered: Why are we and the entire world speaking about it now – virtually 5 months after it was launched?

Effectively, there are 3 primary causes for this:

1. Want for AI Brokers: Since final yr, there was a wave of AI brokers available in the market. Though these brokers have develop into a lot smarter, they nonetheless battle to work effectively with real-world information and instruments (like calendars, emails, or databases). MCP fixes this by standardizing how AI connects to those instruments, making it simpler for them to work collectively.
2. Neighborhood Momentum: Well-liked firms like Cursor, Proceed, Supply Graph, and extra are MCP purchasers, whereas firms like Github, One Drive, Google Drive, and so on. are within the record of its servers. Each these lists are rising daily, as an increasing number of firms be a part of the MCP world.
3. Mannequin-Agnostic & Open Supply: MCP is open-source and works with any AI mannequin. Whether or not you’re utilizing Claude, GPT-4, or some other AI system, MCP connects them to real-world information and instruments. In contrast to closed methods, you aren’t locked into one specific mannequin or service.

These are a couple of of many causes on why the world is speaking about MCP and why it will make waves within the coming few months. 

Conclusion

Mannequin Context Protocol (MCP) is revolutionizing how AI methods work together with the actual world by offering seamless integration with instruments, databases, and native information. With MCP, AI brokers can now effortlessly entry related information and carry out advanced duties without having customized code or handbook integrations. Whether or not it’s bettering customer support, automating workflows, or enhancing private AI assistants, MCP unlocks highly effective new potentialities for AI fashions, making them extra environment friendly, intuitive, and adaptable. As companies and builders proceed to undertake MCP, its potential to boost AI-driven methods will solely develop, opening up new alternatives throughout industries.

Incessantly Requested Questions

Q1. What’s Mannequin Context Protocol (MCP)? 

A. MCP is an open commonplace that permits AI fashions to entry real-world information, equivalent to native information, paperwork, and databases, seamlessly and effectively.

Q2. How does MCP work? 

A. MCP connects an AI shopper (like Claude or an IDE) with a server (native or distant information supply), permitting the AI to request and use context with out handbook enter or customized coding.

Q3. Do I would like a paid subscription for Claude to make use of MCP? 

A. Whereas the free model of Claude works, it typically hits charge limits, making it much less handy. A paid model removes these restrictions.

This fall. What are the essential necessities to make use of MCP? 

A. You’ll want Python 3.11+, Claude Desktop App, an IDE like VS Code, and a few native paperwork for the duties.

Q5. Can MCP work with any AI mannequin? 

A. Sure, MCP is model-agnostic and can be utilized with any AI mannequin, together with Claude, GPT-4, and open-source LLMs.

Q6. What instruments will be linked utilizing MCP? 

A. MCP permits AI to entry a wide range of instruments equivalent to Google Drive, GitHub, Notion, emails, and native information.

Q7. How do I arrange the configuration file for Claude? 

A. It’s essential to go to Claude’s settings below the developer part and create a claude_desktop_config.json file that maps your instruments to Claude’s interface.




Anu Madan has 5+ years of expertise in content material creation and administration. Having labored as a content material creator, reviewer, and supervisor, she has created a number of programs and blogs. At the moment, she engaged on creating and strategizing the content material curation and design round Generative AI and different upcoming expertise.